NVIDIA, ARM, Tegra 3, and Beyond
2009 was a year that NVIDIA would rather forget. AMD beat it to the punch with DX11 parts, the Fermi family of parts were showing many issues, GPGPU was not a growth industry as of yet, and the company was in a costly legal battle with Intel combined with the shutting down of NVIDIA’s previously impressive chipset division. The company had announced the first Tegra processors, but the pickup on those parts was limited almost entirely to the Microsoft Zune HD line. It was not a success.
Tegra 2 is tiny. NVIDIA hopes to expand the design to make it as fast, if not faster, than modern x86 processors.
In 2010 the fortunes of the company changed. While Fermi was first introduced with the hot and power hungry GTX 480, it soon became available as the much more refined and reasonably priced/powered/cooled GTX 460. Seemingly overnight, the fortunes of the company changed dramatically. NVIDIA was looking at a share price that was at a near historic low (around $5 at the very worst), with a stream of successful launches and a change in the company’s direction, their fortunes took a turn for the better. NVIDIA released an entire lineup of DX11 products from the GTS 450 up to the new GTX 580 in the second half of 2010.
Perhaps more significant were the first sightings of the Tegra 2 processor. NVIDIA had licensed the ARM processor architecture for the first Tegra, which was a single core ARM 11 based processor. The ARM 11 was not terribly exciting from an architectural standpoint, but it was a start for NVIDIA. Getting familiar with new CPU designs, integrated with NVIDIA graphics and I/O technologies (integrated southbridge to enable a System on a Chip- SoC), is not an easy task. NVIDIA likely did not expect to conquer this new market overnight, and they likely figured out that the original Tegra will not overtake the low power CPU world by storm. Tegra 2 on the other hand is a totally different ballgame. Tegra 2 is based on the latest generation Cortex A9 design, which is much more powerful and scalable than the previous generations of ARM CPU families.
News of the Tegra 2 perked up many an ear. Not only was this dual core Cortex A9 processor running at 1 GHz, but it also featured an NVIDIA designed graphics processor integrated on the die itself. NVIDIA has been fairly mum on the topic of the exact details of the graphics portion, but from all indications this appears to be a pretty hefty part. Hefty that is, considering it is an integrated graphics processor that runs on a chip smaller than a dime, with a power envelope of less than 1 watt. Estimates range from between 10% to 30% faster than the comparable Imagination Technologies part which is licensed by Texas Instruments for their OMAP 4000 series of chips and significantly faster than the Adreno graphics included with the Qualcom Snapdragon processors- formerly the Imageon graphics division acquired by Qualcom from AMD.
NVIDIA certainly wants to tap into the potential growth of ARM processor shipments.
At CES 2011 the world finally got its first real look at the Tegra 2, and it was featured in a host of products ranging from cell phones to tablets. Labeled as “The World’s First Superphone”, the LG Optimus 2X impressed with its responsiveness, rich featureset, and the robust graphics performance in not just 2D and 3D settings, but also in delivering smooth 1080P video playback when the phone was hooked to a HD TV through a HDMI cable. While the original Tegra cannot be viewed as a commercial success, the amount of products utilizing the Tegra 2 processor is truly impressive. This should be a significant portion of revenue throughout 2011.
NVIDIA is not standing still with Tegra 2 though. The first update should be in production shortly, with introductory products within the next 6 months. The Tegra 2 3D ups the ante by increasing the clockspeed of its dual core A9 based CPU to 1.2 GHz, but it will also support 3D panel technology. Much like the 3D screens that will be introduced with the Nintendo 3DS which do not require glasses, other handheld devices will start to utilize this technology. NVIDIA is offering the Tegra 2 3D as an option for those wishing to be able to push 3D content with as much performance as possible. Again, from all indications, the NVIDIA based Ultra Low Power GeForce GPU featured in Tegra 2 appears to be class leading in terms of performance and featureset. This will still be a TSMC 40 nm based product, but should still be around the same power envelope as the previous Tegra 2. This could well be due to NVIDIA’s increasing familiarity of the Cortex A9 architecture, as well as how the individual products can be redesigned to better complement TSMC’s 40 nm process.
We like the gory details.
The big news now is the leak that NVIDIA will be offering a Tegra 3 later this year. When ARM designed the Cortex A9 architecture, it could be implemented in a single, dual, or quad core design scaling up to 2.5 GHz. Of course, a 40/45 nm quad core at 2.5 GHz would certainly not make it in a sub 1 watt power envelope, but the design is able to scale to those numbers if die size and power consumption were not an overriding issue. There will be at least two primary Tegra 3 designs, one for higher performance tablets, and the other for handheld/cellphone use. The higher end part will be a quad core Cortex A9 based CPU going up to 1.5 GHz in speed. NVIDIA also is implementing a new graphics unit, which promises around 3X the performance of the current Tegra 2. The handheld looks like it is in a bit of flux, but it could come in two flavors. One would be a low power quad core, while the other would be a higher clocked dual core. Both will feature the new GPU.
These products will sample by Q4 2011, and it is generally assumed that these parts will be based on TSMC’s 28 nm process. NVIDIA might go with GLOBALFOUNDRIES at some point with this design, but all indications point to TSMC’s process being more mature and actually available in the timeframe mentioned. GLOBALFOUNDRIES does have a working relationship with ARM, and eventually it could be the go-to foundry when producing high performance ARM based parts. But for the moment, it certainly seems like TSMC actually has a working product.